Trust in God, But Lock Your Car

From therundown.studio contributor Candyce Kelshall, Adjunct Professor University of Buckingham Center of Security and Intelligence Studies. President of the Canadian Association for Security and Intelligence Studies. CEO Proactive Analytics.

Large Language Models

The recursive nature of AI means that eventually knowledge will be “synthetic”

Synthetic Knowledge - The end result of recursive AI guessing/stating what should be reality.

Knowledge/questions - Generate answers. Those answers become the basis of future answers. Are they right or wrong?

How is AI using LLM? These know the rules of language and statistically add words associated with other words to generate sentences.

Chat GPT

Chat GPT works because there are rules to language. It was those language rules and sentiment analysis to generate answers.

It does not mean its right, in fact potentially information will become different based on where you live and the trends of what people ask or type.

2018 data is used and everything after that is generated from what answers people get when they ask LL to answer a question.

Answers determine what future answers become info.

Info is degraded over time recurrent neural network.

Neural Networks

Deep neural networks or deep learning means 150 hidden layers.

NN have ⅔ hidden layers where algorithms “if then what?” answers THEN runs again.

Also it often reveals things we don’t know are there, like some things that are not factored or considered. It identifies those.

Machine Learning

Machine Learning: You tell it what features to look for and identify and clarify.

Deep Learning: It tells you what features of images are common to other images in a dataset.

Learning = because we may not have identified those commonalities or likelihoods.

Deep Learning Target

Target could determine when a woman was pregnant based on purchases.

Purchases of large quantities of unscented lotion occurred around beginning of second trimester.

Purchases of supplements like calcium, magnesium and zinc during first 20 weeks.

Sudden purchase of scent-free soap, extra big bags of cotton balls, hand sanitizer and washcloths meant delivery date was close.

Deep learning could detect pregnancy and delivery date with high levels of accuracy.

Critical Thinking

Critical thinking disappears because they believe the information and trust the AI -driven information as absolute.

Data Poisoning

Data poisoning is part of information warfare

False data deliberately introduced to skew results

Mis/disinformation data

Adversarial Machine Learning

How can we exploit or degrade info given to the public using AI

Poor coding in an LLM leaves hacking vulnerabilities when you don’t fully understand what you are using.

Reliability of Datasets

AI is only as good as the dictionaries and datasets used.

Example: When Urban Dictionary was put into the Watson supercomputer, it read the urban dictionary and started using swear words and slurs because it could not distinguish between profanity and polite language.

The only solution was to delete it manually from the supercomputer.

Facial Recognition Issues

AI facial recognition failures rate with black women 18-30 were 40% higher than with white males. So if the training data is flawed or biases, it will be the norm.